I've had bugs the agents can't fix or figure out. Sometimes those involve third-party, proprietary, broken code (read: Windows APIs). Sometimes they just involve complex deployment situation on the client code (I work on desktop apps) where the agent can't figure out what's wrong/makes wrong assumptions/goes nowhere. Sometimes the agent is just very dumb and tunnels vision on the wrong fix.

FWIW, I've also had bugs the agent fixed that I probably never would've figured out without LLMs - LLMs are definitely useful! But I need to keep understanding how the code works so I can take over the reigns when the LLM fails.